andynoodles/omnidoc-ocr-correction-bench

---
dataset_info:
features:
- name: prompt
dtype: string
- name: image
dtype: image
num_examples: 1355
license: cc-by-4.0
task_categories:
- image-to-text
tags:
- ocr
- document-understanding
- markdown
- paddleocr
- omnidocbench
size_categories:
- 1K<n<10K
---

# OmniDoc OCR Correction Bench

A benchmark dataset for evaluating VLMs on OCR error correction and document-to-markdown formatting.

## Overview

Each sample pairs a document image from [OmniDocBench v1.5](https://github.com/opendatalab/OmniDocBench) with a prompt containing PaddleOCR-extracted markdown text. The task is to correct OCR errors and restore proper formatting using the source image as reference.

## Dataset Structure

| Field | Type | Description |
|-------|------|-------------|
| `prompt` | `string` | System prompt with OCR-extracted markdown text to correct |
| `image` | `image` | Source document image |

## Sources

- **Images**: [OmniDocBench v1.5](https://github.com/opendatalab/OmniDocBench) — covers books, papers, exams, newspapers, magazines, PPTs, notes, textbooks, and financial reports
- **OCR extraction**: PaddleOCR with markdown output
- **Prompts**: Custom correction prompts instructing the model to fix OCR errors while preserving document structure